Dynamic demand forecasting in bike-sharing systems: a multidimensional spatiotemporal network

Accurate demand forecasting is critical for optimising resource allocation and operational efficiency in bike-sharing systems (BSS). However, this remains challenging owing to the complex nonlinear characteristics of large-scale data, irregular spatiotemporal patterns, and external environmental uncertainties. To address these issues, this study proposes a multidimensional spatiotemporal fusion framework that combines Domain-Adaptive Density Clustering (DADC) and an Attention Temporal Graph Convolutional Network (A3T-GCN). First, the DADC algorithm addresses data sparsity by adaptively clustering stations based on spatial density variations, thereby effectively capturing dynamic distribution patterns. Second, a weighted directed graph is constructed incorporating station utility and full-load-rate functions to eliminate low-performance nodes. A dynamic graph sequence model is then developed to characterise the temporal evolution of spatial dependencies and enhance peak-period responsiveness. The A3T-GCN architecture integrates station-specific attributes with environmental factors through attention mechanisms, enabling the focused analysis of critical demand influencers while processing multidimensional spatiotemporal dependencies. Experimental validation using Citi Bike datasets demonstrates superior performance over benchmark models, reducing prediction errors by 11.54% (15 min) and 12.33% (60 min) through improved short-term accuracy and long-term stability. The framework effectively addresses the uncertainty in demand forecasting, providing methodological support for sustainable BSS operations through enhanced spatiotemporal pattern recognition and environmental adaptability.